OverviewWhat it is
An agent is an LLM given tools (search, code, APIs) and the autonomy to decide which to call, observe the result, and loop until a goal is met. This is the shift from answering questions to getting things done. The common pattern is ReAct: interleave reasoning, tool calls, and observations.
At a glanceAgents & Tool Use
An agent decides which tools to call and loops on results. Autonomy adds power - and reliability risk you must contain.
CompareThe agent-pattern landscape, side by side
The patterns you actually compose an agent from — reasoning loops, multi-agent teams, ways to act, and controls. Filter by kind or search across autonomy, cost, and what each is for. Most real agents combine several: a loop, some tools, a guard.
Every row has a page — what it does, what it costs you, and how to tell when it is the thing biting you.
MovedThis grew its own section
The Agent Skills catalogue used to sit on this page. It has its own section now — 31 dated entries, filterable by provider, with a full page each.
MechanicsHow it works
Function/tool calling lets the model emit a structured request the app executes; the result feeds back in. Multi-agent systems split work across specialised agents. Protocols like MCP standardise how models connect to tools and data.
Ground levelWhat you actually build
The loop is four steps. Everything that decides whether it survives contact with production is in the lanes above and below it.
LandscapeTypes & approaches
Click a highlighted type to open its own page — concept, use case, and diagram.
FeasibilityArchitecture & feasibility
Architecture & feasibility
- Feasibility is dominated by reliability and cost: each loop can compound errors and tokens. You need step limits, retries, timeouts, and observability.
- Tool permissioning and least-privilege access are security-critical once an agent can act on real systems.
- State/memory and idempotent tools decide whether an agent can recover from a failed step - a core design concern.
In practiceWhat it means for building
Agents unlock automation but add unpredictability. The PM question is 'where is autonomy worth the reliability risk?' - start with narrow, reversible tasks.
You own guardrails, tool permissions, retries, memory/state, and observability. Error handling and cost control matter more as autonomy grows.
GlossaryKey terms
CheckCheck your understanding
What makes agents hard in production?
Reliability and cost. Loops compound errors and tokens. You need guardrails, permissioning, step limits, retries, and observability.
What is ReAct?
A pattern where the model alternates reasoning and tool actions until it can answer, making its steps inspectable.
How do you keep an agent safe?
Least-privilege tools, human approval for high-stakes actions, step/cost limits, and full tracing.
What changedWhat changed here
As of 2026-09-25 — tool-protocol mentions in the daily brief: 2 items in the last 7 days
Updated this page Agentic features are reaching mobile surfaces, so an agent's integration target can be a phone app rather than a desktop or API.
Updated this page Google opened early access to an MCP server for Google Home, letting agents control devices and query activity, which is a concrete example of a platform exposing a tool surface to third-party agents.
Updated this page A paper shows LLM agents running controlled experiments with simulation models, extending agent use from text and code to understanding system interventions.
Updated this page The paper argues a governed, standardised coding-agent harness drives enterprise agent performance more than model choice or orchestration layers, a prioritisation signal for teams building agents.
Updated this page A survey of multimodal agentic frameworks gives readers a single reference for how perception, memory, and decision-making are orchestrated around LLM backbones.
Three kinds of claim, strongest first. Signal runs every morning.